Start here

Introduction

Tether connects AI to the capabilities your product already has, while keeping production under your control. It is headless: nothing renders, and you keep your interface.

The mental model

Your application posts a question and a tenant identity. Tether decides what the question needs, gathers context, calls a model with your tools, and returns structured JSON along with the exact cost of the turn.

When the model wants to read something, Tether calls an HTTPS endpoint you control and feeds the result back. When the model wants to change something, Tether does not call you at all — it stages the change and returns a preview. Your endpoint sees it only after a person confirms.

Your application
      │
      ▼
    Tether
      │
      ├── Identify intent        no model call
      ├── Check policy           quota, then spend
      ├── Gather context         cache, memory, retrieval
      ├── Call model             with your tool catalog
      ├── Invoke tools           signed request to your endpoint
      ├── Stage writes           nothing executes
      └── Record execution       tokens, dollars, telemetry
      │
      ▼
Your systems

Tether never receives your database credentials. It reaches your data only through the endpoint you expose, and it can only call the tools you registered.

What Tether handles


Routing	Deciding what a question needs before spending a token on it
Model execution	The call, the tool loop, retries, and the turn ceiling
Tool governance	Which tools exist, which are visible on which plan, which write
Staged writes	Preview, confirmation, single execution under a race
Structured output	A typed block contract with repair and safe degradation
Spend	A per-turn ledger, per-tenant budgets, and a breaker that reads them
Memory	Durable user preferences and retrieval over your own documentation
Telemetry	Why a turn routed the way it did, and where the latency went

What Tether is not

Not a chat UI. Nothing renders. You receive typed blocks and build the surface. See Rendering blocks.

Not an agent framework. You do not write graphs, chains, or orchestration code. You register JSON Schema and expose one endpoint.

Not a model provider. Tether orchestrates one. It is not one.

Not a data store for your business records. Tether keeps conversations, staged actions, spend, telemetry, and memory. Your orders, customers, and inventory stay where they are.

When not to use Tether

Be honest with yourself about these — the product is a poor fit for all three.

You want a chatbot on a marketing site. Tether is built for operating a product, not for answering questions about one. The governance is overhead you would not use.
You have no API and no service layer. Tether calls capabilities you already have. If exposing a tool means writing the feature first, the feature is the work; Tether is not.
Nothing you would let AI do carries any consequence. If every operation is a read and nothing costs money or changes state, a framework will serve you at lower cost.

What is not built yet

Where to go next

Quickstart — from nothing to a first answer
Your first write — the part that makes Tether worth using
The tool endpoint — the one thing you implement